2009 09 Ben Cohen SBPL

download 2009 09 Ben Cohen SBPL

of 19

Transcript of 2009 09 Ben Cohen SBPL

  • 8/14/2019 2009 09 Ben Cohen SBPL

    1/19

    My Summer at Willow Garage

    Benjamin CohenUniversity of Pennsylvania

  • 8/14/2019 2009 09 Ben Cohen SBPL

    2/19

    The Bosses

    Dr. Sachin Chitta

    my mentor

    Professor Maxim Likhachev

    my advisor

  • 8/14/2019 2009 09 Ben Cohen SBPL

    3/19

    My Summer Goals

    SBPL Arm Planner

    continue development of new algorithm

    integrate into move_arm

    perform tabletop manipulation

    publish

    SBPL Door Planner(Sachin did all of the integration. I'm working on the algorithm and

    debugging it with him)continue development of algorithm

    make a sweet video

    publish

  • 8/14/2019 2009 09 Ben Cohen SBPL

    4/19

    SBPL Arm Planner

    Goal:

    Develop a motion planner for robotic manipulation thatcan robustly plan paths in cluttered environments.

  • 8/14/2019 2009 09 Ben Cohen SBPL

    5/19

    Why not run IK and plan to a knownvalid joint configuration?

    IK is very fast but may not find the 'best' solution

    may find a visually awkward solution

    may orient the arm very close to obstacles

    may choose a solution very close to joint limits

  • 8/14/2019 2009 09 Ben Cohen SBPL

    6/19

    Why not run IK and plan to a knownvalid joint configuration? (cont'd)

    Let the planner optimize the goal joint configurationaccording to the desired cost function:

    length of path

    smoothnessdistance from obstacles

    keep away from joint limits

  • 8/14/2019 2009 09 Ben Cohen SBPL

    7/19

    Graph Construction

    similar to lattice based planners in planningdynamically feasible trajectories

    a discretization of the configuration space into aset of states, the edges that connect thestarts are dynamically feasible

    edges are pre-defined motion primitives

  • 8/14/2019 2009 09 Ben Cohen SBPL

    8/19

  • 8/14/2019 2009 09 Ben Cohen SBPL

    9/19

    Costs

    Smoothing Cost

    minimize change in velocities between states

    keep the same joint(s) moving if possible

    Action Cost

    minimize the number of actions to the goal

    Distance Cost

    apply a cost for planning close to obstacles

    c s , s ' =ccells 'w

    actionc

    actions , s ' w

    smoothc

    smooths , s '

  • 8/14/2019 2009 09 Ben Cohen SBPL

    10/19

    Heuristic

    3D Dijkstra search

    robust to obstacles

    pre-computed

    directs the end effector to the goal pose

    difference in orientation of the end effector

    Goal

    Dijkstra Shortest Path

  • 8/14/2019 2009 09 Ben Cohen SBPL

    11/19

    ARA : Anytime Random A*

    is consistent

    guarantees completeness

    'anytime results' can trade-off path cost vs.computation time

  • 8/14/2019 2009 09 Ben Cohen SBPL

    12/19

    ARA : Anytime Random A*

    ARA* finds a feasible solution quickly and thencontinually works on improving it until theallocated time runs out.

    ARA* uses an admissible heuristic (a bound onthe plan-execution cost) to focus its search

    Each solution comes with a bound, , on its sub-optimality:

    cost(solution) cost(optimal solution)

  • 8/14/2019 2009 09 Ben Cohen SBPL

    13/19

    Features

    given multiple goals, find the path to goal posewith lowest cost

    plan to 'goal region'

    plan with path constraints, such as:

    keep a cup upright

    don't obstruct field of view of forearm cameras (futurework)

    plan to a goal pose in collision, capable ofmanipulation (future work)

  • 8/14/2019 2009 09 Ben Cohen SBPL

    14/19

    Drawbacks

    high dependency on the heuristic function

    local minima

    doesn't describe kinematics of the arm

    high dimensionality slows down the search

    discretized paths

  • 8/14/2019 2009 09 Ben Cohen SBPL

    15/19

    Future Work

    learn a set of motion primitives with goodcoverage of the workspace and natural lookingmovement

    add randomization to escape local minimaR* is an A* type search with randomization

    research smoothing methods

  • 8/14/2019 2009 09 Ben Cohen SBPL

    16/19

    SBPL Door Planner

    Goal:

    open doors

    inward and outward opening doors

    right handed and left handed doors

    with collision avoidance

    guaranteed feasibility

    Method:

    ARA* in a 4-Dimensional state space{x,y,theta,door interval}

    door interval: {0,1} (open or closed)

  • 8/14/2019 2009 09 Ben Cohen SBPL

    17/19

    Other Packages

    Robot Voxelizer (sandbox/robot_voxelizer)

    wrapper around Ioan's planning models

    voxelize robot bodies, meshes, shapes

    PM Wrapper (planning_research/pm_wrapper)

    wrapper around Ioan's planning_monitor class

    easy collision checking, collision map manipulation

  • 8/14/2019 2009 09 Ben Cohen SBPL

    18/19

    Other Packages (cont'd)

    Arm Command Tool (sandbox/move_arm_tools)

    'safe' movement with move_arm

    'unsafe' movement for tabletop manipulation

  • 8/14/2019 2009 09 Ben Cohen SBPL

    19/19

    Thank You!

    Scott & Steve

    Sachin

    MaxIoan and Mrinal

    Everyone else!

    thanks!